Using multiple points for video prediction

[Kidleyh]

Hello, when I use multiple points for video prediction, I find that there are always a lot of noisy masks in the first frame, but not in the following frames. Is there any way to improve this situation?

At the same time, the video does not seem to support the application of returning a mask (mulit mask output) for each point?